Mapping the spatial distribution of winter crops at sub-pixel level using AVHRR NDVI time series and neural nets

Abstract

For large areas, it is difficult to assess the spatial distribution and inter-annualvariation of crop acreages through field surveys. Such information, however, is of greatvalue for governments, land managers, planning authorities, commodity traders andenvironmental scientists. Time series of coarse resolution imagery offer the advantage ofglobal coverage at low costs, and are therefore suitable for large-scale crop type mapping.Due to their coarse spatial resolution, however, the problem of mixed pixels has to beaddressed. Traditional hard classification approaches cannot be applied because ofsub-pixel heterogeneity. We evaluate neural networks as a modeling tool for sub-pixel cropacreage estimation. The proposed methodology is based on the assumption that differentcover type proportions within coarse pixels prompt changes in time profiles of remotelysensed vegetation indices like the Normalized Difference Vegetation Index (NDVI).Neural networks can learn the relation between temporal NDVI signatures and the soughtcrop acreage information. This learning step permits a non-linear unmixing of the temporalinformation provided by coarse resolution satellite sensors. For assessing the feasibilityand accuracy of the approach, a study region in central Italy (Tuscany) was selected. Thetask consisted of mapping the spatial distribution of winter crops abundances within 1 kmAVHRR pixels between 1988 and 2001. Reference crop acreage information for networktraining and validation was derived from high resolution Thematic Mapper/EnhancedThematic Mapper (TM/ETM+) images and official agricultural statistics. Encouragingresults were obtained demonstrating the potential of the proposed approach. For example, thespatial distribution of winter crop acreage at sub-pixel level was mapped with a cross-validatedcoefficient of determination of 0.8 with respect to the reference information from highresolution imagery. For the eight years for which reference information was available, theroot mean squared error (RMSE) of winter crop acreage at sub-pixel level was 10%. Whencombined with current and future sensors, such as MODIS and Sentinel-3, the unmixing ofAVHRR data can help in the building of an extended time series of crop distributions andcropping patterns dating back to the 80s. © 2013 by the authors; licensee MDPI, Basel, Switzerland.